Minute-level intelligent scheduling method and system for hydrogen-based energy prepared from renewable energy

Abstract

The invention provides a minute-level intelligent scheduling method and system for hydrogen-based energy prepared from renewable energy, and relates to the technical field of energy management and intelligent scheduling, and the method comprises the steps: obtaining data of power generation equipment, hydrogen production equipment and a hydrogen storage system, adaptively adjusting the length of a dynamic time window based on the power generation fluctuation frequency, and generating a fluctuation feature sequence; and constructing a minute-level intelligent scheduling model, optimizing a hydrogen production power regulation instruction by utilizing a mutual feedback correction parameter, calculating an optimal power transfer path when an abnormality is detected, and executing stepped load reduction turn-off. The operation stability of the hydrogen production system can be improved, efficient utilization of generated power is achieved, and the operation risk of the system is reduced.

Description

Technical Field

[0001] The present invention relates to energy management and intelligent scheduling technology, and in particular to a minute-level intelligent scheduling method and system for producing hydrogen-based energy from renewable energy. Background Art

[0002] With the rapid development of renewable energy hydrogen production technology, how to effectively address the intermittent and volatile nature of renewable energy generation and achieve stable operation of hydrogen energy systems has become a critical issue that the industry urgently needs to address. Current renewable energy hydrogen system scheduling methods primarily rely on fixed time windows for power regulation, which is difficult to adapt to the random fluctuations in generated power and can easily lead to delayed system response and insufficient regulation accuracy.

[0003] Existing hydrogen production power regulation schemes often control water electrolysis hydrogen production equipment and hydrogen storage systems as independent units, lacking a system-level collaborative optimization mechanism. This results in low equipment utilization and difficulty achieving a smooth transition in the event of a system failure. Furthermore, due to the complex nature of renewable energy generation, traditional power regulation methods struggle to accurately capture the dynamic characteristics of power generation fluctuations, impacting the overall operational efficiency of the system.

[0004] Traditional scheduling and control methods employ simple emergency shutdown strategies when system anomalies occur. This can easily lead to excessive equipment stress and fluctuating hydrogen production, reducing system reliability and economic efficiency. Therefore, there is an urgent need to develop a minute-by-minute intelligent scheduling method that can adaptively adjust time windows, achieve coordinated optimization of hydrogen storage systems and hydrogen production equipment, and provide smooth fault transition capabilities. Summary of the Invention

[0005] The embodiments of the present invention provide a minute-level intelligent scheduling method and system for producing hydrogen-based energy from renewable energy, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a minute-level intelligent scheduling method for renewable energy hydrogen-based energy, including:

[0008] Obtain real-time power generation data from renewable energy power generation equipment, operating parameter data from water electrolysis hydrogen production equipment, and hydrogen storage capacity data from hydrogen storage systems;

[0009] Calculate the power generation fluctuation frequency based on real-time power generation data, adaptively adjust the length of the dynamic time window based on the power generation fluctuation frequency, and use the dynamic time window to segment the power generation data to generate a fluctuation feature sequence;

[0010] A minute-level intelligent scheduling model is constructed based on the fluctuation characteristic sequence, operating parameter data, and hydrogen storage data to generate the initial hydrogen production power adjustment instruction. The mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment are calculated. The mutual feedback correction parameters are used to compensate and optimize the initial hydrogen production power adjustment instruction to obtain the optimized hydrogen production power adjustment instruction.

[0011] The optimized hydrogen production power adjustment instruction is sent to the control system of the water electrolysis hydrogen production equipment to control the operation of the water electrolysis hydrogen production equipment and monitor the operating status of the hydrogen production system. When an abnormality is detected, the optimal power transfer path is calculated according to the fluctuation characteristic sequence, and the generated power is transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform a step-by-step load reduction and shutdown.

[0012] In an optional embodiment,

[0013] The power generation fluctuation frequency is calculated based on real-time power generation data. The length of the dynamic time window is adaptively adjusted based on the power generation fluctuation frequency. The power generation data is segmented using the dynamic time window to generate a fluctuation feature sequence including:

[0014] Performing wavelet transform on real-time power generation data to obtain a frequency component energy spectrum, calculating the energy value of each frequency interval based on the frequency component energy spectrum, and determining the frequency interval with an energy value greater than a preset energy threshold as the dominant fluctuation frequency interval;

[0015] Calculating a frequency characteristic coefficient based on the energy distribution within the dominant fluctuation frequency interval, inputting the frequency characteristic coefficient into a pre-constructed dynamic time window adaptive model, optimizing model parameters of the dynamic time window adaptive model using a multi-objective optimization algorithm, and calculating the dynamic time window length based on the optimized model parameters;

[0016] Performing sliding window segmentation on the real-time power generation data based on the dynamic time window length to obtain a time window power sequence, and calculating the mean of the time window power sequence to obtain a power mean sequence;

[0017] The standard deviation of the power mean sequence is calculated to obtain the power fluctuation amplitude characteristic, the zero crossing rate of the power mean sequence is calculated to obtain the power fluctuation frequency characteristic, and the rate of change of the power mean sequence is calculated to obtain the power fluctuation change characteristic;

[0018] The power fluctuation amplitude feature, the power fluctuation frequency feature and the power fluctuation change feature are combined to construct a fluctuation feature sequence.

[0019] In an optional embodiment,

[0020] Segmenting the real-time power generation data into sliding windows based on the dynamic time window length includes:

[0021] Performing sliding segmentation processing on real-time power generation data based on the length of a dynamic time window, calculating the power change rate of adjacent sampling points, constructing a data fluctuation assessment function based on the power change rate, evaluating the data integrity of the time window using the data fluctuation assessment function, and determining the time window whose data integrity is lower than a preset integrity threshold as a window to be supplemented;

[0022] Performing feature extraction on the power generation data of the window to be supplemented to obtain a feature vector including a power mean and a fluctuation period, constructing a multi-layer similarity matching rule based on the feature vector, and using the multi-layer similarity matching rule to screen out data segments from the historical power generation data whose similarity with the power generation characteristics of the window to be supplemented is greater than a similarity threshold to form a candidate supplement sequence;

[0023] Setting an initial weight coefficient of a dynamic time warping algorithm according to a power change rate within the window to be supplemented, calculating similarities between each data segment in a candidate supplementary sequence and the window to be supplemented based on the initial weight coefficient, and selecting the data segment with the highest similarity as the supplementary data;

[0024] constructing a weight attenuation function based on the power change rate, and performing weighted smoothing processing on the original power generation data and the supplementary data in the window to be supplemented according to the weight attenuation function to generate supplemented time window data;

[0025] Calculate the sequence continuity index and dynamic consistency index of the supplemented time window data, compare the weighted results of the continuity index and consistency index with the preset evaluation threshold, and when it is less than the preset evaluation threshold, adjust the parameters of the weight decay function and return to execute the weighted smoothing processing step until it is greater than the preset evaluation threshold, and output the time window power sequence that meets the requirements.

[0026] In an optional embodiment,

[0027] The mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment are calculated, and the initial hydrogen production power adjustment instructions are compensated and optimized using the mutual feedback correction parameters. The optimized hydrogen production power adjustment instructions include:

[0028] Based on the pressure value in the hydrogen storage data, a mapping model between pressure and electrolysis efficiency is established, and the deviation of the current pressure from the rated pressure is calculated to obtain the pressure influence factor. Based on the temperature value in the operating parameter data, a mapping model between temperature and hydrogen production efficiency is established, and the deviation of the current temperature from the rated temperature is calculated to obtain the temperature coupling factor. The pressure influence factor is multiplied by the temperature coupling factor, and combined with the equipment performance attenuation value to obtain the mutual feedback correction parameter.

[0029] The initial hydrogen production power regulation instruction is compensated and adjusted according to the mutual feedback correction parameter to obtain a corrected power instruction; at the same time, the maximum power limit of the electrolyzer is compensated to obtain a corrected power upper limit, and the power climbing rate limit of the electrolyzer is compensated to obtain a corrected climbing rate limit;

[0030] A smoothing factor is determined based on the time rate of change of the mutual feed correction parameter, and the corrected power instruction and the power regulation instruction of the previous control cycle are weightedly averaged according to the smoothing factor to generate a transition instruction; it is judged whether the transition instruction meets the corrected power upper limit and the corrected climbing rate limit. If so, the transition instruction is output as the optimized hydrogen production power regulation instruction; if not, the transition instruction is limited within an allowable range and then output as the optimized hydrogen production power regulation instruction.

[0031] In an optional embodiment,

[0032] Monitor the operating status of the hydrogen production system. When an anomaly is detected, calculate the optimal power transfer path based on the fluctuation characteristic sequence, transfer the generated power to the grid according to the optimal power transfer path, and control the water electrolysis hydrogen production equipment to perform a step-by-step load reduction shutdown, including:

[0033] The membrane voltage fluctuation rate, hydrogen production efficiency decay rate, and system pressure mutation rate of the water electrolysis hydrogen production equipment are collected to calculate the first-order rate of change and the second-order rate of change; when the first-order rate of change or the second-order rate of change exceeds the corresponding warning threshold, the power transfer mechanism is triggered;

[0034] Obtaining a power generation sequence of a water electrolysis hydrogen production device and extracting power fluctuation characteristics; combining the first-order rate of change, the second-order rate of change, and the power fluctuation characteristics, and using a deep reinforcement learning algorithm to generate an initial power transfer path;

[0035] Collect stress monitoring data of the water electrolysis hydrogen production equipment and calculate a stress state index based on the stress monitoring data; divide the load reduction process into a buffer load reduction section, a rapid load reduction section, and a protection load reduction section according to the stress state index; establish a dynamic mapping relationship between the stress state index and the load reduction rate of each load reduction section to generate a step-by-step load reduction strategy;

[0036] Performing wavelet decomposition on the pressure signal of the hydrogen storage system to extract the disturbance component and the trend component; identifying the pressure pulsation characteristics based on the disturbance component, and predicting the pressure change trend based on the trend component; combining the pressure pulsation characteristics and the pressure change trend with the stress state index to obtain the system evolution characteristics;

[0037] The system evolution characteristics are used to perform real-time correction on the initial power transfer path to obtain an optimal power transfer path; the generated power is controlled to be transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform load reduction based on the step-by-step load reduction strategy;

[0038] The stress monitoring data and pressure signal feedback during the load reduction process are used to calculate the updated system evolution characteristics; based on the updated system evolution characteristics, the optimal power transfer path and the load reduction rate in the step-by-step load reduction strategy are optimized in real time; and the optimization process is repeated until the water electrolysis hydrogen production equipment is safely shut down.

[0039] In an optional embodiment,

[0040] Obtaining the power generation sequence of the water electrolysis hydrogen production equipment and extracting the power fluctuation characteristics; combining the first-order change rate, second-order change rate and power fluctuation characteristics, using a deep reinforcement learning algorithm to generate an initial power transfer path includes:

[0041] Input power fluctuation characteristics into the strategy network and value network;

[0042] The first hidden layer of the strategy network uses an exponential linear unit function to perform nonlinear combination on the power fluctuation characteristics to obtain a power feature vector, the second hidden layer processes the power feature vector through a residual connection to obtain a power adjustment action, and the output layer generates an action probability of power transfer based on the power adjustment action;

[0043] The long-term and short-term memory units of the value network store power fluctuation characteristics to form a historical power sequence, and the historical power sequence is screened for importance through a gating mechanism to obtain a state assessment result;

[0044] Calculating a power change acceleration coefficient, a power fluctuation amplitude coefficient, and a power response time coefficient according to the power fluctuation characteristics, constructing a power fluctuation constraint rule based on the power change acceleration coefficient, the power fluctuation amplitude coefficient, and the power response time coefficient, and combining the power fluctuation constraint rule with the action probability of the power transfer to form a multi-objective optimization function;

[0045] The reward value is calculated using the multi-objective optimization function and the state evaluation result, a temporal difference error is constructed according to the reward value, a trust domain constraint is established based on the temporal difference error, the parameters of the policy network are optimized under the trust domain constraint, and an initial power transfer path is generated using the optimized policy network parameters.

[0046] In an optional embodiment,

[0047] The load shedding process is divided into a buffer load shedding section, a rapid load shedding section, and a protection load shedding section according to the stress state index; a dynamic mapping relationship between the stress state index and the load shedding rate of each load shedding section is established to generate a step-by-step load shedding strategy, including:

[0048] Establishing a state characteristic matrix based on the stress state index and calculating a characteristic root combination of the state characteristic matrix; determining the system stability state based on the characteristic root combination and calculating a stability margin based on the system stability state; dividing the load shedding process into a buffer load shedding section, a rapid load shedding section, and a protection load shedding section based on the stability margin;

[0049] Establishing a response function between load shedding rate and dynamic stress for each load shedding section; determining the initial load shedding rates of the buffer load shedding section, the rapid load shedding section, and the protection load shedding section based on the response function;

[0050] A dynamic equation of stress state index and load reduction rate is established, wherein a nonlinear term is introduced to characterize the bifurcation characteristics of the system; the initial load reduction rate is substituted into the dynamic equation to solve the bifurcation parameter;

[0051] Establishing an adaptive adjustment equation for stress deviation; calculating a compensation amount based on the deviation between the stress state indicator and a preset stress threshold; inputting the compensation amount into the adaptive adjustment equation to dynamically adjust the bifurcation parameter; solving the dynamic equation based on the adjusted bifurcation parameter to obtain the load shedding rate of each load shedding section;

[0052] A neural network model with multiple hidden layers is constructed, historical load shedding data is input into the neural network model for training, a mapping relationship between load shedding rate and stress fluctuation is established using the trained neural network model, a load shedding rate correction value is calculated based on the mapping relationship, and the load shedding rate of each load shedding section is corrected in combination with the load shedding rate correction value;

[0053] A load shedding time allocation function is established, and the switching moment and duration of each load shedding segment are calculated according to the time allocation function; and the switching moment, duration and the corrected load shedding rate are combined to obtain a step-by-step load shedding strategy.

[0054] According to a second aspect of the embodiments of the present invention,

[0055] A system is provided, comprising:

[0056] The first unit is used to obtain real-time power generation data of renewable energy power generation equipment, operating parameter data of water electrolysis hydrogen production equipment, and hydrogen storage capacity data of the hydrogen storage system;

[0057] The second unit is used to calculate the power generation fluctuation frequency based on the real-time power generation data, adaptively adjust the length of the dynamic time window based on the power generation fluctuation frequency, and use the dynamic time window to segment the power generation data to generate a fluctuation feature sequence;

[0058] The third unit is used to build a minute-level intelligent scheduling model based on the fluctuation characteristic sequence, operating parameter data and hydrogen storage data, generate the initial hydrogen production power adjustment instruction, calculate the mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment, and use the mutual feedback correction parameters to compensate and optimize the initial hydrogen production power adjustment instruction to obtain the optimized hydrogen production power adjustment instruction;

[0059] The fourth unit is used to send the optimized hydrogen production power adjustment instruction to the control system of the water electrolysis hydrogen production equipment, control the operation of the water electrolysis hydrogen production equipment, and monitor the operating status of the hydrogen production system. When an abnormality is detected, the optimal power transfer path is calculated according to the fluctuation characteristic sequence, the generated power is transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform a step-by-step load reduction and shutdown.

[0060] According to a third aspect of the embodiments of the present invention,

[0061] An electronic device is provided, comprising:

[0062] processor;

[0063] a memory for storing processor-executable instructions;

[0064] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0065] According to a fourth aspect of the embodiments of the present invention,

[0066] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0067] In this embodiment, by adaptively adjusting the length of the dynamic time window and generating a fluctuation feature sequence, the fluctuation characteristics of renewable energy power generation can be accurately captured, the ability to identify and predict power generation fluctuations is improved, and the hydrogen production system can respond to power generation fluctuations more flexibly, thereby improving the stability and reliability of the system. The initial hydrogen production power adjustment instruction is compensated and optimized using mutual feedback correction parameters, which realizes the coordinated operation between the hydrogen storage system and the water electrolysis hydrogen production equipment, effectively solves the energy waste problem caused by parameter mismatch in traditional scheduling methods, and improves the energy utilization efficiency and economy of the system. An exception handling mechanism based on the optimal power transfer path is designed. When an abnormality occurs in the system, the generated power can be intelligently transferred to the power grid and a step-by-step load reduction shutdown is performed, avoiding equipment damage and system collapse risks, while minimizing energy losses and enhancing the safety and resilience of the entire hydrogen production system. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1This is a flow chart of a minute-level intelligent scheduling method for producing hydrogen-based energy from renewable energy according to an embodiment of the present invention;

[0069] Figure 2 The reliability and stability evaluation analysis diagram of the supplementary data for the embodiment of the present invention;

[0070] Figure 3 This is a comparison chart of system energy efficiency of an embodiment of the present invention;

[0071] Figure 4 This is a simulation diagram of the dynamic adjustment of the load shedding rate in each stage in the step-by-step load shedding strategy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0073] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0074] Figure 1 FIG. 1 is a flow chart of a minute-level intelligent scheduling method for producing hydrogen-based energy from renewable energy according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0075] Obtain real-time power generation data from renewable energy power generation equipment, operating parameter data from water electrolysis hydrogen production equipment, and hydrogen storage capacity data from hydrogen storage systems;

[0076] Calculate the power generation fluctuation frequency based on real-time power generation data, adaptively adjust the length of the dynamic time window based on the power generation fluctuation frequency, and use the dynamic time window to segment the power generation data to generate a fluctuation feature sequence;

[0077] A minute-level intelligent scheduling model is constructed based on the fluctuation characteristic sequence, operating parameter data, and hydrogen storage data to generate the initial hydrogen production power adjustment instruction. The mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment are calculated. The mutual feedback correction parameters are used to compensate and optimize the initial hydrogen production power adjustment instruction to obtain the optimized hydrogen production power adjustment instruction.

[0078] The optimized hydrogen production power adjustment instruction is sent to the control system of the water electrolysis hydrogen production equipment to control the operation of the water electrolysis hydrogen production equipment and monitor the operating status of the hydrogen production system. When an abnormality is detected, the optimal power transfer path is calculated according to the fluctuation characteristic sequence, and the generated power is transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform a step-by-step load reduction and shutdown.

[0079] In an optional embodiment,

[0080] The power generation fluctuation frequency is calculated based on real-time power generation data. The length of the dynamic time window is adaptively adjusted based on the power generation fluctuation frequency. The power generation data is segmented using the dynamic time window to generate a fluctuation feature sequence including:

[0081] Performing wavelet transform on real-time power generation data to obtain a frequency component energy spectrum, calculating the energy value of each frequency interval based on the frequency component energy spectrum, and determining the frequency interval with an energy value greater than a preset energy threshold as the dominant fluctuation frequency interval;

[0082] Calculating a frequency characteristic coefficient based on the energy distribution within the dominant fluctuation frequency interval, inputting the frequency characteristic coefficient into a pre-constructed dynamic time window adaptive model, optimizing model parameters of the dynamic time window adaptive model using a multi-objective optimization algorithm, and calculating the dynamic time window length based on the optimized model parameters;

[0083] Performing sliding window segmentation on the real-time power generation data based on the dynamic time window length to obtain a time window power sequence, and calculating the mean of the time window power sequence to obtain a power mean sequence;

[0084] The standard deviation of the power mean sequence is calculated to obtain the power fluctuation amplitude characteristic, the zero crossing rate of the power mean sequence is calculated to obtain the power fluctuation frequency characteristic, and the rate of change of the power mean sequence is calculated to obtain the power fluctuation change characteristic;

[0085] The power fluctuation amplitude feature, the power fluctuation frequency feature and the power fluctuation change feature are combined to construct a fluctuation feature sequence.

[0086] For example, real-time power generation data is subjected to a wavelet transform to obtain the frequency component energy spectrum. Specifically, a multi-resolution wavelet decomposition method is used, selecting the db4 wavelet basis to perform a five-layer decomposition of the power generation data, obtaining wavelet coefficients for different frequency intervals. The energy value for each frequency interval is obtained by calculating the square sum of the wavelet coefficients at each layer. For example, for power generation data with a sampling frequency of 1 Hz, after a five-layer wavelet decomposition, energy values are obtained for the five frequency intervals of 0-0.03 Hz, 0.03-0.06 Hz, 0.06-0.12 Hz, 0.12-0.25 Hz, and 0.25-0.5 Hz.

[0087] A preset energy threshold of 15% of the total energy is set. When the energy value of a frequency interval exceeds this threshold, it is identified as the dominant fluctuation frequency interval. For example, if the energy value of the 0.06-0.12Hz frequency interval accounts for 25% and the energy value of the 0.12-0.25Hz frequency interval accounts for 20%, these two frequency intervals are determined to be the dominant fluctuation frequency intervals.

[0088] The frequency characteristic coefficient is calculated based on the energy distribution within the dominant fluctuation frequency range. Specifically, the frequency characteristic coefficient is obtained by dividing the sum of the product of the center frequency of the dominant fluctuation frequency range and its energy value by the total energy. For example, if the dominant fluctuation frequency ranges are 0.06-0.12Hz and 0.12-0.25Hz, their center frequencies are 0.09Hz and 0.185Hz respectively, and the energy values are 250 units and 200 units respectively, and the total energy is 1000 units, then the frequency characteristic coefficient is (0.09×250+0.185×200) / 1000=0.06.

[0089] Based on the calculated frequency characteristic coefficient, the dynamic time window adaptive model is used to determine the time window length. The model adopts a BP neural network structure, with the input layer being the frequency characteristic coefficient, the hidden layer containing 10 neurons, and the output layer being the time window length. Optionally, the model parameters are optimized using a particle swarm optimization algorithm, with the optimization objectives including the adaptability of the window length and computational efficiency. The particle swarm size is set to 50, the maximum number of iterations is 100, the inertia weight decreases linearly from 0.9 to 0.4, and the acceleration factor is 2.0. For example, when the frequency characteristic coefficient is 0.06, the time window length calculated by the optimized model is 180 seconds.

[0090] The real-time power generation data is segmented into sliding windows using the calculated dynamic time window length. The window sliding step is set to 1 / 4 of the window length, or 45 seconds for a 180-second window length. The arithmetic mean of the power data within each time window is calculated to form a power mean sequence. For example, for a photovoltaic power plant's power generation data, if the power mean within the first time window (0-180 seconds) is 500 kW, and the power mean within the second time window (45-225 seconds) is 520 kW, and so on, the resulting power mean sequence is [500, 520, 535, 510, 490, ...] kW.

[0091] Based on the power mean sequence, three fluctuation characteristics are calculated. First, the power fluctuation amplitude characteristic, i.e., the standard deviation of the power mean sequence, is calculated. Subtract the mean from the power mean sequence, calculate the sum of squared deviations, divide by the sequence length, and take the square root to obtain the standard deviation. For example, for the power mean sequence [500, 520, 535, 510, 490, ...] kW, if its standard deviation is 15 kW, then the power fluctuation amplitude characteristic is 15 kW.

[0092] Next, calculate the power fluctuation frequency characteristic, namely the zero-crossing rate of the power mean sequence. Subtract the mean from the power mean sequence, count the number of times the sequence value changes from negative to positive or vice versa, and divide by the sequence length to obtain the zero-crossing rate. For example, for the de-averaged power sequence [-11, 9, 24, -1, -21, ...] kW, if there are 12 zero-crossings in a 100-point sequence, the zero-crossing rate is 0.12.

[0093] Finally, calculate the power fluctuation characteristic, that is, the rate of change of the power mean sequence. Calculate the difference between adjacent power mean values and average their absolute values to obtain the average rate of change. For example, for the power mean sequence [500, 520, 535, 510, 490, ...] kW, the absolute values of the differences between adjacent points are [20, 15, 25, 20, ...] kW. If the average value is 18 kW, then the power fluctuation characteristic is 18 kW.

[0094] The three features above are combined to construct a fluctuation feature sequence. For each time window, the power fluctuation amplitude feature, power fluctuation frequency feature, and power fluctuation change feature are combined into a three-dimensional feature vector. If the power fluctuation amplitude feature of a time window is 15kW, the power fluctuation frequency feature is 0.12, and the power fluctuation change feature is 18kW, then the fluctuation feature vector for this window is [15, 0.12, 18]. As the time window slides, a series of fluctuation feature vectors are obtained, forming a complete fluctuation feature sequence.

[0095] In this embodiment, efficient utilization of renewable energy generation is achieved, improving the responsiveness of the water electrolysis hydrogen production system and adapting it to the real-time fluctuations of renewable energy generation. By calculating the power generation fluctuation frequency and dynamically adjusting the time window, short-term changes in generated power can be accurately captured, making scheduling decisions more flexible and improving the system's adaptability to fluctuating power sources. A minute-level intelligent scheduling model is employed to fine-tune hydrogen production power regulation, reducing energy waste and system instability caused by lags in traditional scheduling methods. By optimizing scheduling instructions through feedback correction parameters, the energy flow between the hydrogen storage system and the hydrogen production equipment can be effectively coordinated, avoiding hydrogen storage system overload or unstable hydrogen production equipment operation, and improving overall operational efficiency. When the system detects an anomaly, the optimal power transfer path is calculated based on the fluctuation characteristic sequence, enabling the timely and efficient distribution of generated power to the grid or other loads, avoiding energy waste or grid impact caused by sudden failures of the hydrogen production equipment. A stepped load reduction and shutdown strategy can gradually reduce the hydrogen production load during power adjustment, reducing equipment losses caused by sudden shutdowns or severe fluctuations, extending equipment life, and ensuring safe and stable system operation.

[0096] In an optional embodiment,

[0097] Segmenting the real-time power generation data into sliding windows based on the dynamic time window length includes:

[0098] Performing sliding segmentation processing on real-time power generation data based on the length of a dynamic time window, calculating the power change rate of adjacent sampling points, constructing a data fluctuation assessment function based on the power change rate, evaluating the data integrity of the time window using the data fluctuation assessment function, and determining the time window whose data integrity is lower than a preset integrity threshold as a window to be supplemented;

[0099] Performing feature extraction on the power generation data of the window to be supplemented to obtain a feature vector including a power mean and a fluctuation period, constructing a multi-layer similarity matching rule based on the feature vector, and using the multi-layer similarity matching rule to screen out data segments from the historical power generation data whose similarity with the power generation characteristics of the window to be supplemented is greater than a similarity threshold to form a candidate supplement sequence;

[0100] Setting an initial weight coefficient of a dynamic time warping algorithm according to a power change rate within the window to be supplemented, calculating similarities between each data segment in a candidate supplementary sequence and the window to be supplemented based on the initial weight coefficient, and selecting the data segment with the highest similarity as the supplementary data;

[0101] constructing a weight attenuation function based on the power change rate, and performing weighted smoothing processing on the original power generation data and the supplementary data in the window to be supplemented according to the weight attenuation function to generate supplemented time window data;

[0102] Calculate the sequence continuity index and dynamic consistency index of the supplemented time window data, compare the weighted results of the continuity index and consistency index with the preset evaluation threshold, and when it is less than the preset evaluation threshold, adjust the parameters of the weight decay function and return to execute the weighted smoothing processing step until it is greater than the preset evaluation threshold, and output the time window power sequence that meets the requirements.

[0103] Exemplarily, the real-time power generation data is first subjected to sliding segmentation processing based on the dynamic time window length. Specifically, the initial time window length is set to 30 minutes and the sliding step is 5 minutes. For the data in each time window, the power change rate of adjacent sampling points is calculated. Assuming that the power value at time t is P(t) and the power value at time t+1 is P(t+1), the power change rate can be expressed as the ratio of the power difference between the two moments to the time interval. For example, if P(t) = 100kW, P(t+1) = 105kW, and the sampling interval is 1 minute, the power change rate is 5kW / min.

[0104] A data fluctuation assessment function is constructed based on the power change rate. This function comprehensively considers the average, maximum, and standard deviation of the power change rate within a window to determine the data completeness index for that time window. For example, within a 30-minute window, the average power change rate is 3 kW / min, the maximum is 8 kW / min, and the standard deviation is 1.5 kW / min. The weighted calculation yields a data completeness of 0.85. If the preset completeness threshold is 0.9, the window is identified as a window to be supplemented.

[0105] Perform feature extraction on the power generation data in the window to be supplemented to obtain a feature vector. The feature vector includes the power mean, standard deviation, fluctuation period, and number of peaks. For example, the feature vector for a window to be supplemented is: power mean 120kW, standard deviation 15kW, main fluctuation period 10 minutes, and number of peaks 3.

[0106] A multi-layer similarity matching rule is constructed based on the feature vectors. This rule first filters by power mean, retaining historical data segments whose power mean differs by no more than 20% from the window to be supplemented. It then filters by fluctuation period, retaining segments whose fluctuation period differs by no more than 30%. Finally, a comprehensive consideration is given to the similarity of standard deviation and number of peaks. Through this multi-layered screening, data segments with a similarity greater than 0.85 (a preset similarity threshold) are selected from the historical database to form candidate supplementary sequences.

[0107] The initial weight coefficient of the dynamic time warping algorithm is set based on the power change rate within the window to be supplemented. If the power change rate within the window is large, indicating severe power generation fluctuations, a higher time elasticity coefficient is set. For example, if the average power change rate is 5 kW / min, the initial weight coefficient is set to 0.7. If the power change rate is small, a lower time elasticity coefficient is set. For example, if the average power change rate is 1 kW / min, the initial weight coefficient is set to 0.3.

[0108] The similarity between each data segment in the candidate supplementary sequence and the window to be supplemented is calculated based on the initial weight coefficient. For example, if there are three data segments in the candidate sequence, A, B, and C, and their similarities with the window to be supplemented are calculated to be 0.92, 0.88, and 0.95, respectively, then data segment C with the highest similarity is selected as the supplementary data.

[0109] A weight decay function is constructed based on the power change rate. This function determines the weight distribution between the original data and the supplementary data during the fusion process. At the edges of the window to be supplemented, the original data is given a higher weight; in the middle, the supplementary data is given a higher weight. For example, for a portion with a high power change rate, the original data might have a weight of 0.3, while the supplementary data might have a weight of 0.7. For a portion with a low power change rate, the original data might have a weight of 0.6, while the supplementary data might have a weight of 0.4.

[0110] The original power generation data and supplementary data within the supplementary window are weighted and smoothed using a weight decay function. For example, at a certain moment, the original data is 95 kW and the supplementary data is 105 kW. The weight of the original data is 0.4, and the weight of the supplementary data is 0.6. The smoothed data is 95 × 0.4 + 105 × 0.6 = 101 kW. This process processes all data points within the window to generate the supplemented time window data.

[0111] Calculate the serial continuity index and dynamic consistency index for the supplemented time window data. The serial continuity index assesses the smoothness of the data before and after supplementation, for example, the calculated continuity index is 0.88. The dynamic consistency index assesses the consistency of the changing trends of the supplemented data with the original data, for example, the calculated consistency index is 0.92. Assuming a weight of 0.4 for the continuity index and 0.6 for the consistency index, the weighted result is 0.88 × 0.4 + 0.92 × 0.6 = 0.904.

[0112] The weighted results of the continuity and consistency indices are compared with a preset evaluation threshold (e.g., 0.9). If the weighted result, 0.904, is greater than the preset evaluation threshold, the power sequence of the current supplemented time window is output. If it is less than the preset evaluation threshold, the parameters of the weight decay function are adjusted. For example, the weight of the original data can be increased by 0.1, and then the weighted smoothing step is repeated until the weighted result exceeds the preset evaluation threshold.

[0113] Through the above method, incomplete windows in power generation data can be effectively identified and intelligently supplemented with historical data to ensure that the output power sequence meets the continuity and consistency requirements, providing reliable data support for subsequent power generation forecasting and grid scheduling.

[0114] Figure 2 The reliability and stability evaluation analysis diagram of the supplementary data of the embodiment of the present invention is as follows: Figure 2 The chart below comprehensively demonstrates the reliability and stability of different supplementation methods under long-term operating conditions. A one-year reliability trend analysis of supplementation data is included. The long-term trend analysis shows that the proposed solution (circles) maintains high supplementation data reliability throughout the year, consistently ranging between 88% and 91%, demonstrating excellent long-term stability. Notably, even during challenging periods of high summer loads (June to August) and low winter temperatures (November to January), its reliability index fluctuates only slightly, reaching a low of 88% in July. In contrast, the reliability index of the sliding window regression method (triangles) fluctuates significantly throughout the year, dropping from 81% in winter and spring to 72% in summer, a fluctuation of 9 percentage points. The multi-model ensemble method (squares) performs moderately well, with a reliability index fluctuating between 79% and 86%.

[0115] The two key time points marked in the chart: "Equipment Upgrade" in July and "Algorithm Optimization" in September, also had varying impacts on each method. After the equipment upgrade, the sliding window regression method's performance declined significantly, indicating poor adaptability to equipment changes. In contrast, our solution demonstrated good adaptability, maintaining stable reliability metrics. After algorithm optimization, the performance of all three methods improved, but our solution achieved the most significant improvement, demonstrating its ability to more effectively leverage the performance gains from algorithm optimization.

[0116] Existing technologies use fixed time windows to segment power generation data, which makes it difficult to adapt to the fluctuating characteristics of renewable energy, resulting in insufficient data integrity and affecting scheduling accuracy. Traditional data completion methods mainly rely on linear interpolation or historical mean filling, ignoring power change trends, which may introduce large errors and reduce data credibility. This application accurately identifies windows with insufficient data integrity by adaptively adjusting the time window length and constructing a data fluctuation assessment function based on the power change rate. Combining feature extraction with multi-layer similarity matching rules, highly similar segments are selected from historical data for supplementation, avoiding the error accumulation caused by traditional interpolation methods. In addition, a dynamic time warping algorithm is introduced to optimize the matching results and improve the accuracy of the supplemented data. A weight decay function is used to weight and smooth the supplemented data and the original data to ensure a natural data transition. The data quality is evaluated through continuity and consistency indicators, and the weight parameters are dynamically adjusted to ensure that the final data meets the integrity and consistency requirements. Compared with traditional methods, this solution more flexibly adapts to the dynamic changes of power generation data, provides more reliable input for intelligent scheduling, and improves the operating efficiency and stability of the hydrogen production system.

[0117] In an optional embodiment,

[0118] The mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment are calculated, and the initial hydrogen production power adjustment instructions are compensated and optimized using the mutual feedback correction parameters. The optimized hydrogen production power adjustment instructions include:

[0119] Based on the pressure value in the hydrogen storage data, a mapping model between pressure and electrolysis efficiency is established, and the deviation of the current pressure from the rated pressure is calculated to obtain the pressure influence factor. Based on the temperature value in the operating parameter data, a mapping model between temperature and hydrogen production efficiency is established, and the deviation of the current temperature from the rated temperature is calculated to obtain the temperature coupling factor. The pressure influence factor is multiplied by the temperature coupling factor, and combined with the equipment performance attenuation value to obtain the mutual feedback correction parameter.

[0120] The initial hydrogen production power regulation instruction is compensated and adjusted according to the mutual feedback correction parameter to obtain a corrected power instruction; at the same time, the maximum power limit of the electrolyzer is compensated to obtain a corrected power upper limit, and the power climbing rate limit of the electrolyzer is compensated to obtain a corrected climbing rate limit;

[0121] A smoothing factor is determined based on the time rate of change of the mutual feed correction parameter, and the corrected power instruction and the power regulation instruction of the previous control cycle are weightedly averaged according to the smoothing factor to generate a transition instruction; it is judged whether the transition instruction meets the corrected power upper limit and the corrected climbing rate limit. If so, the transition instruction is output as the optimized hydrogen production power regulation instruction; if not, the transition instruction is limited within an allowable range and then output as the optimized hydrogen production power regulation instruction.

[0122] This embodiment provides a method for calculating the mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment, and uses the parameters to compensate and optimize the initial hydrogen production power adjustment instructions.

[0123] Specifically, a mapping model of pressure and electrolysis efficiency is first established based on the pressure value in the hydrogen storage data. In practical applications, the electrolysis efficiency of water electrolysis hydrogen production equipment will change with the change of the pressure of the hydrogen storage system. By collecting real-time pressure data of the hydrogen storage system, the deviation of the current pressure relative to the rated pressure can be calculated. For example, when the rated pressure of the hydrogen storage system is 30MPa, and the current measured pressure is 25MPa, the pressure deviation is -5MPa. According to the pre-established pressure-efficiency mapping relationship, the pressure influence factor corresponding to the pressure deviation can be determined. Specifically, this can be achieved by looking up a table. For example, when the pressure deviation is -5MPa, the corresponding pressure influence factor is 1.05, indicating that the electrolysis efficiency is increased by 5% when the pressure is lower than the rated pressure.

[0124] Based on the temperature values in the operating parameter data, a mapping model of temperature and hydrogen production efficiency is established. In the process of hydrogen production by electrolysis of water, temperature is a key factor affecting the efficiency of hydrogen production. By collecting the real-time temperature data of the electrolyzer, the deviation of the current temperature relative to the rated temperature is calculated. For example, when the rated operating temperature of the electrolyzer is 75°C and the current measured temperature is 65°C, the temperature deviation is -10°C. According to the pre-established temperature-efficiency mapping relationship, the temperature coupling factor corresponding to the temperature deviation can be determined. For example, when the temperature deviation is -10°C, the corresponding temperature coupling factor is 0.92, which means that the hydrogen production efficiency is reduced by 8% below the rated temperature.

[0125] The pressure influence factor is multiplied by the temperature coupling factor and combined with the equipment performance degradation value to obtain the mutual feedback correction parameter. The equipment performance degradation value reflects the performance degradation of the electrolyzer over time. For example, assuming the electrolyzer has been in operation for 5000 hours, the equipment performance degradation value obtained based on historical data analysis is 0.98. The mutual feedback correction parameter is then calculated as: 1.05 × 0.92 × 0.98 = 0.947, indicating that after comprehensively considering pressure, temperature, and equipment degradation, the actual hydrogen production capacity is 94.7% of the theoretical value.

[0126] The initial hydrogen production power control command is compensated and adjusted based on the mutual feedback correction parameter to obtain a corrected power command. Assuming the initial power control command is 1000kW, the corrected power command is 1000 ÷ 0.947 = 1056kW, indicating that the input power needs to be increased to achieve the expected hydrogen production.

[0127] Compensate for the electrolyzer's maximum power limit to obtain a corrected power upper limit. For example, if the electrolyzer's rated maximum power is 1500kW, the corrected power upper limit is 1500 × 0.947 = 1420.5kW, indicating that the electrolyzer's actual maximum power should be reduced under current operating conditions. Compensate for the electrolyzer's power ramp rate limit to obtain a corrected ramp rate limit. For example, if the electrolyzer's standard ramp rate limit is 10kW / s, the corrected ramp rate limit is 10 × 0.947 = 9.47kW / s, indicating that the power change rate should be reduced under current conditions to protect the equipment.

[0128] The smoothing factor is determined based on the time rate of change of the mutual-feed correction parameter. Changes in the mutual-feed correction parameter reflect fluctuations in the system's operating state. To avoid frequent changes in the power command, a smoothing mechanism is introduced. Specifically, the rate of change of the current mutual-feed correction parameter compared to the previous cycle's mutual-feed correction parameter is calculated. If the rate of change is large, the smoothing factor is smaller; if the rate of change is small, the smoothing factor is larger. For example, when the rate of change of the mutual-feed correction parameter is 2%, the corresponding smoothing factor is 0.7.

[0129] The corrected power command and the power regulation command from the previous control cycle are weighted averaged according to the smoothing factor to generate the transition command. Assuming the power regulation command from the previous cycle was 980 kW, the current corrected power command is 1056 kW, and the smoothing factor is 0.7, the transition command is calculated as: 1056 × 0.7 + 980 × 0.3 = 1033.2 kW.

[0130] Finally, determine whether the transition instruction meets the corrected power upper limit and corrected ramp rate limit. If so, the transition instruction is output as the optimized hydrogen production power adjustment instruction; if not, the transition instruction is limited to the allowable range before output. For example, if the actual power executed in the previous cycle is 1000kW, the current transition instruction is 1033.2kW, the corrected ramp rate limit is 9.47kW / s, and the control cycle is 2s, then the maximum power change allowed in a single cycle is 9.47×2=18.94kW. Since 1033.2-1000=33.2kW>18.94kW, which exceeds the ramp rate limit, the final optimized power adjustment instruction output should be 1000+18.94=1018.94kW.

[0131] In this embodiment, by establishing a mapping model between pressure and electrolysis efficiency, and temperature and hydrogen production efficiency, the pressure influence factor and temperature coupling factor are accurately calculated, and combined with the equipment performance attenuation value, the mutual feedback correction parameter is obtained, so that the power adjustment instruction can dynamically adapt to the actual operating state and avoid energy efficiency degradation caused by environmental changes. By compensating and adjusting the initial hydrogen production power adjustment instruction through the mutual feedback correction parameter, the accuracy of the power instruction can be optimized. At the same time, the maximum power limit and power ramp rate of the electrolyzer are compensated to ensure that the power adjustment process meets the equipment operation safety requirements, fully utilizes renewable energy within a reasonable range, and reduces equipment losses caused by power fluctuations. Based on the time rate of change of the mutual feedback correction parameter, a smoothing factor is introduced to make power regulation smoother and avoid the impact of sudden adjustment on the equipment. Finally, by judging whether the transition instruction meets the corrected power upper limit and ramp rate limit, it is ensured that the optimized power adjustment instruction not only meets the operating requirements of the hydrogen storage system and hydrogen production equipment, but also can be dynamically adjusted within a safe range, thereby improving the safety, response speed and energy utilization efficiency of the overall system.

[0132] In an optional embodiment,

[0133] Monitor the operating status of the hydrogen production system. When an anomaly is detected, calculate the optimal power transfer path based on the fluctuation characteristic sequence, transfer the generated power to the grid according to the optimal power transfer path, and control the water electrolysis hydrogen production equipment to perform a step-by-step load reduction shutdown, including:

[0134] The membrane voltage fluctuation rate, hydrogen production efficiency decay rate, and system pressure mutation rate of the water electrolysis hydrogen production equipment are collected to calculate the first-order rate of change and the second-order rate of change; when the first-order rate of change or the second-order rate of change exceeds the corresponding warning threshold, the power transfer mechanism is triggered;

[0135] Obtaining a power generation sequence of a water electrolysis hydrogen production device and extracting power fluctuation characteristics; combining the first-order rate of change, the second-order rate of change, and the power fluctuation characteristics, and using a deep reinforcement learning algorithm to generate an initial power transfer path;

[0136] Collect stress monitoring data of the water electrolysis hydrogen production equipment and calculate a stress state index based on the stress monitoring data; divide the load reduction process into a buffer load reduction section, a rapid load reduction section, and a protection load reduction section according to the stress state index; establish a dynamic mapping relationship between the stress state index and the load reduction rate of each load reduction section to generate a step-by-step load reduction strategy;

[0137] Performing wavelet decomposition on the pressure signal of the hydrogen storage system to extract the disturbance component and the trend component; identifying the pressure pulsation characteristics based on the disturbance component, and predicting the pressure change trend based on the trend component; combining the pressure pulsation characteristics and the pressure change trend with the stress state index to obtain the system evolution characteristics;

[0138] The system evolution characteristics are used to perform real-time correction on the initial power transfer path to obtain an optimal power transfer path; the generated power is controlled to be transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform load reduction based on the step-by-step load reduction strategy;

[0139] The stress monitoring data and pressure signal feedback during the load reduction process are used to calculate the updated system evolution characteristics; based on the updated system evolution characteristics, the optimal power transfer path and the load reduction rate in the step-by-step load reduction strategy are optimized in real time; and the optimization process is repeated until the water electrolysis hydrogen production equipment is safely shut down.

[0140] For example, a sensor network collects key operating parameters of a water electrolysis hydrogen production system in real time, including membrane voltage, hydrogen production flow rate, and system pressure. Based on this collected data, the membrane voltage fluctuation rate, hydrogen production efficiency decay rate, and system pressure mutation rate are calculated. Specifically, the membrane voltage fluctuation rate is calculated by dividing the membrane voltage difference between adjacent time points by the time interval; the hydrogen production efficiency decay rate is calculated by the rate of change of hydrogen production per unit power; and the system pressure mutation rate is calculated by the rate of change of the pressure sensor data.

[0141] Taking the membrane voltage fluctuation rate as an example, the system collects membrane voltage data every 100 milliseconds and calculates the first-order change rate, that is, the change rate of two adjacent sampling points; at the same time, it calculates the second-order change rate, that is, the speed of change of the first-order change rate. When the first-order change rate of membrane voltage exceeds 0.5V / s or the second-order change rate exceeds 0.8V / s 2 When the system is judged as an abnormal state, the power transfer mechanism is triggered.

[0142] After triggering the power transfer mechanism, the system acquires the generated power sequence of the water electrolysis hydrogen production equipment, typically consisting of power data points at 5-second intervals over the last 30 minutes. Time-frequency analysis is used to extract power fluctuation characteristics, including amplitude, frequency distribution, and fluctuation trends. For example, during one abnormal situation, the system recorded a power fluctuation amplitude of 8% of the rated power, with the main fluctuation frequency concentrated in the 0.05-0.1Hz range and showing a gradually increasing trend.

[0143] The system takes first-order rate of change, second-order rate of change, and power fluctuation characteristics as input and generates an initial power transfer path using a pre-trained deep reinforcement learning model. This model utilizes a two-layer LSTM network architecture, with the input layer containing the feature dimension, the hidden layer containing 128 nodes, and the output layer representing the power transfer sequence. Model training uses historical anomaly handling data, with the reward function being to minimize system stress during power transfers and maximize grid stability. The generated initial power transfer path includes a time series from current power to zero power. For example, in one anomaly handling event, the initial path was to smoothly transfer 10 MW of power to the grid within 120 seconds.

[0144] Simultaneously, the system collects stress monitoring data from the water electrolysis hydrogen production equipment, including stack temperature gradients, membrane stress distribution, and differential pressure across the hydrogen-oxygen separator. Based on this data, a comprehensive stress state index is calculated, ranging from 0 to 100, with 0 representing a stress-free state and 100 representing an extreme stress state. The load shedding process is divided into three phases based on the stress state index: a buffer load shedding phase (stress index 0-30), with a load shedding rate of 1-3% / second of rated power; a rapid load shedding phase (stress index 31-70), with a load shedding rate of 3-8% / second of rated power; and a protective load shedding phase (stress index 71-100), with a load shedding rate reduced to 0.5-1% / second of rated power, simultaneously activating the auxiliary cooling system.

[0145] The system performs wavelet decomposition on the hydrogen storage system's pressure signal, using the db4 wavelet basis function for a five-layer decomposition to extract high-frequency disturbance components and low-frequency trend components. Based on the disturbance component, the system identifies pressure pulsation characteristics, such as frequency, amplitude, and phase. Based on the trend component, it predicts the pressure change trend over the next 30 seconds. For example, during an anomaly handling incident, the system identified a pressure pulsation frequency of 1.2 Hz and an amplitude of 0.3 MPa, with the trend component indicating a pressure increase of 0.05 MPa / s.

[0146] The system combines pressure pulsation characteristics, pressure change trends, and stress state indicators to form a system evolution feature vector. This feature vector is input into the adaptive correction module, which performs real-time corrections on the initial power transfer path to obtain the optimal power transfer path. This correction process takes into account the grid's capacity, electrolyzer safety limitations, and the pressure state of the hydrogen storage system. The transfer path is updated every 5 seconds using a rolling horizon optimization method.

[0147] Based on the optimal power transfer path, the system controls the transfer of generated power to the grid through power electronic conversion equipment, while simultaneously controlling the electrolysis of water to hydrogen production equipment to perform load shedding based on a step-by-step load shedding strategy. In specific implementations, power transfer and load shedding are carried out in tandem. For example, in one process, the system first reduces power by 20% at a rate of 2% / second (buffer load shedding), then reduces power by 50% at a rate of 5% / second (rapid load shedding), and finally reduces the remaining 30% at a rate of 0.8% / second (protection load shedding).

[0148] During the load shedding process, the system continuously collects stress monitoring data and pressure signals, calculating updated system evolution characteristics every two seconds. Based on these updated characteristics, the system optimizes the power transfer path and load shedding rate in real time. For example, if a sudden increase in the stack temperature gradient is detected, the system adjusts the load shedding rate during the rapid load shedding phase from 5% / second to 4% / second and extends the protective load shedding phase.

[0149] The system repeated the optimization process until the water electrolysis hydrogen production equipment was completely shut down. Throughout the process, the system recorded key parameter curves, including power transfer curves, stress state change curves, and pressure change curves, for subsequent analysis and model optimization. In actual application, this method successfully handled multiple abnormal situations, shortening abnormality handling time by an average of 35% and reducing peak equipment stress by 42%, effectively improving the safety and reliability of the hydrogen production system.

[0150] In this embodiment, by collecting the membrane voltage fluctuation rate, hydrogen production efficiency decay rate and system pressure mutation rate, the first-order and second-order change rates are calculated to accurately identify system anomalies and trigger the power transfer mechanism. The initial power transfer path is generated in combination with the deep reinforcement learning algorithm to make power scheduling more adaptive. At the same time, a dynamic load reduction mapping relationship is constructed based on stress monitoring data, and the load reduction process is divided into buffering, fast and protection load reduction segments to achieve more refined load reduction control and avoid equipment damage or excessive system fluctuations. Wavelet decomposition technology is used to extract the pressure disturbance and trend components of the hydrogen storage system. Combined with the stress state index, the system evolution characteristics are constructed, and the power transfer path and load reduction strategy are continuously optimized based on real-time feedback to ensure that the load reduction rate adapts to the current system state. It can achieve accurate response to abnormal conditions, optimize power transfer, improve system safety and stability, and at the same time ensure the long-term operation reliability of water electrolysis hydrogen production equipment.

[0151] In an optional embodiment,

[0152] Obtaining the power generation sequence of the water electrolysis hydrogen production equipment and extracting the power fluctuation characteristics; combining the first-order change rate, second-order change rate and power fluctuation characteristics, using a deep reinforcement learning algorithm to generate an initial power transfer path includes:

[0153] Input power fluctuation characteristics into the strategy network and value network;

[0154] The first hidden layer of the strategy network uses an exponential linear unit function to perform nonlinear combination on the power fluctuation characteristics to obtain a power feature vector, the second hidden layer processes the power feature vector through a residual connection to obtain a power adjustment action, and the output layer generates an action probability of power transfer based on the power adjustment action;

[0155] The long-term and short-term memory units of the value network store power fluctuation characteristics to form a historical power sequence, and the historical power sequence is screened for importance through a gating mechanism to obtain a state assessment result;

[0156] Calculating a power change acceleration coefficient, a power fluctuation amplitude coefficient, and a power response time coefficient according to the power fluctuation characteristics, constructing a power fluctuation constraint rule based on the power change acceleration coefficient, the power fluctuation amplitude coefficient, and the power response time coefficient, and combining the power fluctuation constraint rule with the action probability of the power transfer to form a multi-objective optimization function;

[0157] The reward value is calculated using the multi-objective optimization function and the state evaluation result, a temporal difference error is constructed according to the reward value, a trust domain constraint is established based on the temporal difference error, the parameters of the policy network are optimized under the trust domain constraint, and an initial power transfer path is generated using the optimized policy network parameters.

[0158] For example, first, the power generation sequence of the electrolytic water hydrogen production device is obtained. In practical applications, the power data can be collected every 5 seconds by the power sensor on the device. The data is continuously collected for 24 hours to obtain a power sequence P(t) consisting of 17280 sampling points. For example, the power data (unit: kW) of a certain electrolytic water hydrogen production device from 8:00 to 8:05 in the morning may be: [120.5, 121.2, 119.8, 122.3, 123.1, 121.7, 120.9, 122.5, 123.8, 124.2, 123.5, 122.8].

[0159] Next, we extract power fluctuation characteristics. These characteristics include the power mean, standard deviation, maximum value, minimum value, crest factor, and coefficient of variation. Taking the power data above as an example, we calculate the power mean to be 122.19 kW, the standard deviation to be 1.37 kW, the maximum value to be 124.2 kW, the minimum value to be 119.8 kW, the crest factor (the ratio of the maximum value to the mean) to be 1.016, and the coefficient of variation (the ratio of the standard deviation to the mean) to be 0.011.

[0160] Then, calculate the first-order rate of change and the second-order rate of change of the power sequence. The first-order rate of change indicates the speed of change of power at adjacent time points, and the second-order rate of change indicates the change in the speed of power change. For the above power data, the first-order rate of change sequence (unit: kW / s) is: [0.14, -0.28, 0.5, 0.16, -0.28, -0.16, 0.32, 0.26, 0.08, -0.14, -0.14], the second-order rate of change sequence (unit: kW / s 2 ) is: [-0.42, 0.78, -0.34, -0.44, 0.12, 0.48, -0.06, -0.18, -0.22, 0].

[0161] Combining the first-order rate of change, second-order rate of change, and power fluctuation characteristics, the specific steps for generating the initial power transfer path using the deep reinforcement learning algorithm are as follows:

[0162] The power fluctuation features are fed into the policy network and the value network. These two networks are core components of deep reinforcement learning. The policy network is responsible for generating actions (i.e., power adjustment decisions), while the value network is responsible for evaluating state values. The input power fluctuation feature vector contains the eigenvalues calculated above and has a dimension of 6.

[0163] The strategy network is designed as follows: The first hidden layer contains 64 neurons and uses an exponential linear unit function to perform a nonlinear combination of power fluctuation characteristics. The exponential linear unit function maintains a linear relationship for positive inputs and exponentially decays for negative inputs, helping to address the vanishing gradient problem. This layer processes the power feature vector, which has a dimension of 64. The second hidden layer, containing 32 neurons, processes the power feature vector using residual connections. Residual connections directly add the input to the output, which helps to address the difficulty of training deep networks. This layer outputs power adjustment actions, which have a dimension of 16. Based on the power adjustment actions, the output layer generates the action probability of power transfer, which has a dimension of 5, corresponding to five possible power adjustment strategies: a large increase (10%), a small increase (5%), no change, a small decrease (5%), and a large decrease (10%).

[0164] The value network uses long-short-term memory (LSTM) units to store power fluctuation characteristics, forming a historical power sequence. The LSTM unit contains 32 memory cells and can store power data for the past 10 time steps. A gating mechanism (including input gates, forget gates, and output gates) filters the historical power sequence for importance, generating a state assessment result. The input gate controls the degree to which new information enters the memory cell, the forget gate controls the degree to which old information is retained, and the output gate controls the degree to which information is output from the memory cell. After processing, the value network outputs a scalar value representing the value assessment of the current state.

[0165] Based on the power fluctuation characteristics, the power change acceleration coefficient, power fluctuation amplitude coefficient, and power response time coefficient are calculated. The power change acceleration coefficient is proportional to the root mean square of the second-order rate of change, and for the above data, it is calculated to be 0.38. The power fluctuation amplitude coefficient is proportional to the power standard deviation, and is calculated to be 0.42. The power response time coefficient is inversely proportional to the mean of the first-order rate of change, and is calculated to be 0.29. Based on these three coefficients, power fluctuation constraint rules are established to ensure that power adjustments do not exceed the physical limitations of the equipment.

[0166] The power fluctuation constraint rules and the power transfer action probability are combined into a multi-objective optimization function. The multi-objective optimization function comprehensively considers the efficiency and safety of power adjustment, with efficiency weighted at 0.6 and safety weighted at 0.4.

[0167] The reward value is calculated using a multi-objective optimization function and state evaluation results. The reward value consists of three parts: a power tracking accuracy reward (weight 0.5), a power smoothness reward (weight 0.3), and an energy efficiency reward (weight 0.2). The temporal difference error (TDE) is constructed based on the reward value. The TDE is the difference between the current state value estimate and the next state value estimate plus the immediate reward.

[0168] A trust region constraint is established based on the temporal difference error to limit the step size of the policy update and prevent excessive policy updates from causing training instability. The trust region radius is set to 0.02, indicating that the KL divergence between the new and old policies does not exceed 0.02. The policy network parameters are optimized under the trust region constraint using the Adam optimizer with a learning rate of 0.001 and 1000 training iterations.

[0169] The optimized policy network parameters generate an initial power transfer path. For the power data above, the generated initial power transfer path is [120.5, 121.2, 120.5, 121.5, 122.6, 122.0, 121.0, 122.0, 123.0, 123.5, 123.0, 122.5]. Compared to the original power sequence, the fluctuation amplitude is reduced by 18% and the power change acceleration is reduced by 25%, while maintaining the average level of power output.

[0170] Figure 3 This is a comparison chart of the energy efficiency of the system according to the embodiment of the present invention. Figure 3The results show the performance of five different control methods in terms of energy efficiency and process parameters of water electrolysis hydrogen production equipment. Under light load conditions (25% load), the proposed technology achieves an efficiency of 72.3%, a 27.3% improvement over the 56.8% achieved with traditional PID control. Under medium load conditions (50% load), the efficiency is 79.8%, a 16.8% improvement. Under heavy load conditions (85% load), the efficiency is 82.5%, a 14.4% improvement. This demonstrates that the proposed technology is particularly suitable for light load conditions and can effectively address the low energy efficiency of traditional water electrolysis hydrogen production equipment under low loads. Regarding electrolytic cell temperature stability, the proposed technology achieves a temperature fluctuation of only ±1.4°C, a 73.6% reduction compared to the ±5.3°C achieved with traditional control. The temperature rise time is shortened from 28.6 minutes to 16.2 minutes, a 43.4% improvement. The maximum temperature rise is only 5.3°C, a 58.6% reduction compared to the 12.8°C achieved with traditional control. This excellent temperature control capability is due to the deep reinforcement learning algorithm's precise control of power changes and comprehensive consideration of temperature influencing factors. In terms of hydrogen purity, the average purity of this technical solution reaches 99.8%, and the minimum purity is not less than 99.3%, which are 1.1% and 2.1% higher than traditional control, respectively. Although the increase in purity seems small, considering that it is already in the high purity range, this improvement is of great significance to hydrogen energy applications, especially fuel cell applications. Overall, this technical solution achieves fine control of the entire working condition of hydrogen production by water electrolysis through the nonlinear combination of the long and short-term memory units of the value network and the strategy network, significantly improving the system energy efficiency and process indicators.

[0171] Existing technologies typically use simple rules or fixed thresholds for power scheduling. However, these methods perform poorly when faced with complex power fluctuations. They often fail to respond promptly to fluctuations, resulting in inaccurate regulation and impacting system stability and efficiency. These traditional methods fail to fully leverage historical data and power fluctuation characteristics to make more dynamic and accurate decisions.

[0172] This application adopts a deep reinforcement learning algorithm, combines the policy network and the value network to model the power fluctuation characteristics, and significantly improves the ability to generate power transfer paths. The policy network nonlinearly combines the power fluctuation characteristics through the exponential linear unit function, and optimizes the power regulation action through the residual connection to make the regulation more precise. The value network uses long and short-term memory units to process historical power sequences, and combines the gating mechanism to filter important information, thereby improving the accuracy and real-time performance of state assessment. In addition, based on the acceleration coefficient, amplitude coefficient and response time coefficient of power fluctuation, this application constructs a power fluctuation constraint rule and combines it with the multi-objective optimization function to ensure that the generated power transfer path not only meets the equipment performance requirements, but also effectively suppresses fluctuations. By optimizing the timing difference error and trust domain constraints, the stability of the policy network is enhanced, making the power regulation process more stable and efficient. These improvements enable this application to intelligently adjust the power transfer path according to real-time power fluctuations and historical data, improve the scheduling capability of the water electrolysis hydrogen production system in a complex dynamic environment, and ensure that the system runs more smoothly and efficiently in fluctuations.

[0173] In an optional embodiment,

[0174] The load shedding process is divided into a buffer load shedding section, a rapid load shedding section, and a protection load shedding section according to the stress state index; a dynamic mapping relationship between the stress state index and the load shedding rate of each load shedding section is established to generate a step-by-step load shedding strategy, including:

[0175] Establishing a state characteristic matrix based on the stress state index and calculating a characteristic root combination of the state characteristic matrix; determining the system stability state based on the characteristic root combination and calculating a stability margin based on the system stability state; dividing the load shedding process into a buffer load shedding section, a rapid load shedding section, and a protection load shedding section based on the stability margin;

[0176] Establishing a response function between load shedding rate and dynamic stress for each load shedding section; determining the initial load shedding rates of the buffer load shedding section, the rapid load shedding section, and the protection load shedding section based on the response function;

[0177] A dynamic equation of stress state index and load reduction rate is established, wherein a nonlinear term is introduced to characterize the bifurcation characteristics of the system; the initial load reduction rate is substituted into the dynamic equation to solve the bifurcation parameter;

[0178] Establishing an adaptive adjustment equation for stress deviation; calculating a compensation amount based on the deviation between the stress state indicator and a preset stress threshold; inputting the compensation amount into the adaptive adjustment equation to dynamically adjust the bifurcation parameter; solving the dynamic equation based on the adjusted bifurcation parameter to obtain the load shedding rate of each load shedding section;

[0179] A neural network model with multiple hidden layers is constructed, historical load shedding data is input into the neural network model for training, a mapping relationship between load shedding rate and stress fluctuation is established using the trained neural network model, a load shedding rate correction value is calculated based on the mapping relationship, and the load shedding rate of each load shedding section is corrected in combination with the load shedding rate correction value;

[0180] A load shedding time allocation function is established, and the switching moment and duration of each load shedding segment are calculated according to the time allocation function; and the switching moment, duration and the corrected load shedding rate are combined to obtain a step-by-step load shedding strategy.

[0181] For example, a state characteristic matrix is first established based on stress state indicators, and load shedding is divided into three stages. The system collects real-time stress state indicators, including grid frequency, voltage, and power. For a specific power system, 10 key indicators, including frequency deviation, voltage deviation, and power margin, are collected to construct a 10×10 state characteristic matrix. By calculating the eigenvalues of this matrix, eigenvalues such as λ1 = 0.82, λ2 = 0.76, and λ3 = 0.65 are obtained. When the real parts of all eigenvalues are negative and the absolute value of the real part of the largest eigenvalue is greater than 0.5, the system is considered stable. Based on the distribution of the eigenvalues, the stability margin is calculated to be 0.68. The load shedding process is divided into three stages based on a preset threshold: when the stability margin is greater than 0.7, it is a buffer load shedding stage; when the stability margin is between 0.4 and 0.7, it is a rapid load shedding stage; and when the stability margin is less than 0.4, it is a protective load shedding stage.

[0182] For each load shedding segment, a response function was established that correlates the load shedding rate with the dynamic stress. Through historical data analysis, the initial load shedding rate for the buffering segment was determined to be 5MW / min, the rapid load shedding segment to be 15MW / min, and the protective load shedding segment to be 3MW / min. For a 500MW unit, for example, in the buffering segment, when the load was reduced by 5MW per minute, the system frequency fluctuation remained within ±0.02Hz; in the rapid load shedding segment, when the load was reduced at a rate of 15MW / min, the frequency fluctuation was controlled within ±0.05Hz; and in the protective load shedding segment, when the load was reduced at a rate of 3MW / min, the frequency fluctuation was controlled within ±0.01Hz.

[0183] Next, a dynamic equation was established that relates the stress state index to the deloading rate. This equation contains both linear and nonlinear terms, with the nonlinear term used to characterize the system's bifurcation characteristics near the critical point. Substituting the initial deloading rate into the dynamic equation, the bifurcation parameters α = 0.62, β = 0.85, and γ = 0.41 were obtained. These parameters describe the critical conditions for the system to transition from one stable state to another.

[0184] To achieve dynamic adjustment of the load shedding rate, an adaptive regulation equation for stress deviation was established. The deviation between the stress state indicator and the preset stress threshold was monitored in real time. When the frequency deviation exceeded 0.1 Hz, a compensation of 0.15 was calculated; when the voltage deviation exceeded 5%, a compensation of 0.2 was calculated. The compensation was input into the adaptive regulation equation, and the bifurcation parameters were dynamically adjusted to α' = 0.71, β' = 0.92, and γ' = 0.38. The dynamic equations were solved based on the adjusted bifurcation parameters, resulting in a load shedding rate adjustment of 6.2 MW / min for the buffer load shedding segment, 17.5 MW / min for the rapid load shedding segment, and 2.8 MW / min for the protective load shedding segment.

[0185] To further optimize the load shedding rate, a neural network model was constructed with three hidden layers, each containing 20, 15, and 10 neurons, respectively. The input layer receives stress state indicators, and the output layer outputs load shedding rate corrections. The model was trained using historical load shedding data, including 100 sets of normal operating conditions and 50 sets of fault conditions. After training, the model achieved an accuracy of 92.5%. This model established a mapping relationship between load shedding rate and stress fluctuations, and calculated load shedding rate corrections: +0.8 MW / min for the buffering load shedding segment, -1.2 MW / min for the rapid load shedding segment, and +0.5 MW / min for the protection load shedding segment. The load shedding rates for each segment were adjusted based on these corrections, resulting in a final load shedding rate of 7.0 MW / min for the buffering segment, 16.3 MW / min for the rapid load shedding segment, and 3.3 MW / min for the protection load shedding segment.

[0186] Finally, a load shedding time allocation function was established to calculate the switching time and duration of each load shedding segment based on the system's initial and target loads. For a system reducing power from 500MW to 300MW, the calculated duration of the buffer load shedding segment was 15 minutes, starting at 0 minutes and ending at 15 minutes; the rapid load shedding segment lasted 10 minutes, starting at 15 minutes and ending at 25 minutes; and the protection load shedding segment lasted 12 minutes, starting at 25 minutes and ending at 37 minutes. Combining the switching time, duration, and corrected load shedding rates yielded a complete step-by-step load shedding strategy: From 0 to 15 minutes, the load shedding rate was 7.0MW / min, reducing power by 105MW; from 15 to 25 minutes, the load shedding rate was 16.3MW / min, reducing power by 163MW; and from 25 to 37 minutes, the load shedding rate was 3.3MW / min, reducing power by 39.6MW. The total load shedding rate was 307.6MW, approaching the target load shedding rate of 300MW, with an acceptable error.

[0187] The stepped load reduction strategy can effectively reduce system oscillations, lower the frequency fluctuation amplitude by about 40%, shorten the system recovery and stability time by about 25%, and significantly improve the safety and stability of the power system.

[0188] Figure 4 This is a simulation diagram of the dynamic adjustment of the load shedding rate in each stage in the step-by-step load shedding strategy according to an embodiment of the present invention. Figure 4 As shown, the chart shows in detail the dynamic rate adjustment process of the three load shedding stages in this technical solution. The buffer load shedding section (0-60 seconds) adopts an average load shedding rate of 0.42MPa / s, and the rate shows a slow downward trend, providing sufficient adaptation time for the system. The load shedding rate in the rapid load shedding section (60-120 seconds) is greatly increased to 1.57MPa / s, reaching a peak at 80 seconds, and then slightly adjusted according to the system response. The load shedding rate in the protection load shedding section (120-180 seconds) is rapidly reduced to 0.25MPa / s and continues to decrease, ensuring that the system smoothly transitions to the final state. The figure shows that the load shedding rate transitions smoothly between each section, avoiding the common mutation problem in traditional methods. Especially at the two turning points of 60 seconds and 120 seconds, the load shedding rate changes smoothly, thanks to the nonlinear bifurcation characteristic analysis and adaptive adjustment algorithm applied in this solution, which effectively prevents the system from oscillations caused by rate mutations.

[0189] Based on the above technical solution, more precise load shedding process control can be achieved, and the load shedding rate can be dynamically adjusted through stress state indicators, thereby improving system stability and the safety of the load shedding process. Existing technologies usually perform load shedding through fixed rates or simple preset modes, but often ignore the real-time changes and dynamic responses of system stress, resulting in the load shedding process being inflexible and difficult to cope with complex operating environments. This application constructs a dynamic mapping relationship between stress state indicators and load shedding rates and uses a multi-level neural network model to train historical data, thereby achieving precise correction of the load shedding rate. By establishing a dynamic equation containing nonlinear terms and introducing an adaptive adjustment mechanism, the load shedding rate can be dynamically adjusted to optimize the load shedding process, especially in the case of large stress fluctuations, which can effectively reduce system instability. In addition, by managing the load shedding process in sections and combining the calculation of stability margins and bifurcation parameters, this solution ensures that the load shedding process can be reasonably adjusted according to different stress states, avoiding the potential risks caused by the single load shedding strategy in existing technologies. The starting point of the improvement is to address the problem of traditional methods ignoring the actual stress and dynamic changes of the system, making the load shedding process more flexible and safe. The improvement significantly enhances the responsiveness and adaptability of the load shedding process, while avoiding the negative impact of overly rough load shedding schemes on the system, thereby improving the safety and stability of the system.

[0190] According to a second aspect of the embodiments of the present invention,

[0191] Provided is a minute-level intelligent scheduling system for hydrogen-based energy produced from renewable energy, the system comprising:

[0192] The first unit is used to obtain real-time power generation data of renewable energy power generation equipment, operating parameter data of water electrolysis hydrogen production equipment, and hydrogen storage capacity data of the hydrogen storage system;

[0193] The second unit is used to calculate the power generation fluctuation frequency based on the real-time power generation data, adaptively adjust the length of the dynamic time window based on the power generation fluctuation frequency, and use the dynamic time window to segment the power generation data to generate a fluctuation feature sequence;

[0194] The third unit is used to build a minute-level intelligent scheduling model based on the fluctuation characteristic sequence, operating parameter data and hydrogen storage data, generate the initial hydrogen production power adjustment instruction, calculate the mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment, and use the mutual feedback correction parameters to compensate and optimize the initial hydrogen production power adjustment instruction to obtain the optimized hydrogen production power adjustment instruction;

[0195] The fourth unit is used to send the optimized hydrogen production power adjustment instruction to the control system of the water electrolysis hydrogen production equipment, control the operation of the water electrolysis hydrogen production equipment, and monitor the operating status of the hydrogen production system. When an abnormality is detected, the optimal power transfer path is calculated according to the fluctuation characteristic sequence, the generated power is transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform a step-by-step load reduction and shutdown.

[0196] According to a third aspect of the embodiments of the present invention,

[0197] An electronic device is provided, comprising:

[0198] processor;

[0199] a memory for storing processor-executable instructions;

[0200] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0201] According to a fourth aspect of the embodiments of the present invention,

[0202] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0203] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A minute-level intelligent scheduling method for renewable energy hydrogen-based energy, characterized in that: include: Obtain real-time power generation data from renewable energy power generation equipment, operating parameter data from water electrolysis hydrogen production equipment, and hydrogen storage capacity data from hydrogen storage systems; Calculate the power generation fluctuation frequency based on real-time power generation data, adaptively adjust the length of the dynamic time window based on the power generation fluctuation frequency, and use the dynamic time window to segment the power generation data to generate a fluctuation feature sequence; A minute-level intelligent scheduling model is constructed based on the fluctuation characteristic sequence, operating parameter data, and hydrogen storage data to generate the initial hydrogen production power adjustment instruction. The mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment are calculated. The mutual feedback correction parameters are used to compensate and optimize the initial hydrogen production power adjustment instruction to obtain the optimized hydrogen production power adjustment instruction. The optimized hydrogen production power adjustment instruction is sent to the control system of the water electrolysis hydrogen production equipment to control the operation of the water electrolysis hydrogen production equipment and monitor the operating status of the hydrogen production system. When an abnormality is detected, the optimal power transfer path is calculated according to the fluctuation characteristic sequence, and the generated power is transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform a step-by-step load reduction and shutdown.

2. The method according to claim 1, characterized in that The power generation fluctuation frequency is calculated based on real-time power generation data. The length of the dynamic time window is adaptively adjusted based on the power generation fluctuation frequency. The power generation data is segmented using the dynamic time window to generate a fluctuation feature sequence including: Performing wavelet transform on real-time power generation data to obtain a frequency component energy spectrum, calculating the energy value of each frequency interval based on the frequency component energy spectrum, and determining the frequency interval with an energy value greater than a preset energy threshold as the dominant fluctuation frequency interval; Calculating a frequency characteristic coefficient based on the energy distribution within the dominant fluctuation frequency interval, inputting the frequency characteristic coefficient into a pre-constructed dynamic time window adaptive model, optimizing model parameters of the dynamic time window adaptive model using a multi-objective optimization algorithm, and calculating the dynamic time window length based on the optimized model parameters; Performing sliding window segmentation on the real-time power generation data based on the dynamic time window length to obtain a time window power sequence, and calculating the mean of the time window power sequence to obtain a power mean sequence; The standard deviation of the power mean sequence is calculated to obtain the power fluctuation amplitude characteristic, the zero crossing rate of the power mean sequence is calculated to obtain the power fluctuation frequency characteristic, and the rate of change of the power mean sequence is calculated to obtain the power fluctuation change characteristic; The power fluctuation amplitude feature, the power fluctuation frequency feature and the power fluctuation change feature are combined to construct a fluctuation feature sequence.

3. The method according to claim 2, characterized in that Segmenting the real-time power generation data into sliding windows based on the dynamic time window length includes: Performing sliding segmentation processing on real-time power generation data based on the length of a dynamic time window, calculating the power change rate of adjacent sampling points, constructing a data fluctuation assessment function based on the power change rate, evaluating the data integrity of the time window using the data fluctuation assessment function, and determining the time window whose data integrity is lower than a preset integrity threshold as a window to be supplemented; Performing feature extraction on the power generation data of the window to be supplemented to obtain a feature vector including a power mean and a fluctuation period, constructing a multi-layer similarity matching rule based on the feature vector, and using the multi-layer similarity matching rule to screen out data segments from the historical power generation data whose similarity with the power generation characteristics of the window to be supplemented is greater than a similarity threshold to form a candidate supplement sequence; Setting an initial weight coefficient of a dynamic time warping algorithm according to a power change rate within the window to be supplemented, calculating similarities between each data segment in a candidate supplementary sequence and the window to be supplemented based on the initial weight coefficient, and selecting the data segment with the highest similarity as the supplementary data; constructing a weight attenuation function based on the power change rate, and performing weighted smoothing processing on the original power generation data and the supplementary data in the window to be supplemented according to the weight attenuation function to generate supplemented time window data; Calculate the sequence continuity index and dynamic consistency index of the supplemented time window data, compare the weighted results of the continuity index and consistency index with the preset evaluation threshold, and when it is less than the preset evaluation threshold, adjust the parameters of the weight decay function and return to execute the weighted smoothing processing step until it is greater than the preset evaluation threshold, and output the time window power sequence that meets the requirements.

4. The method according to claim 1, wherein The mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment are calculated, and the initial hydrogen production power adjustment instructions are compensated and optimized using the mutual feedback correction parameters. The optimized hydrogen production power adjustment instructions include: Based on the pressure value in the hydrogen storage data, a mapping model between pressure and electrolysis efficiency is established, and the deviation of the current pressure from the rated pressure is calculated to obtain the pressure influence factor. Based on the temperature value in the operating parameter data, a mapping model between temperature and hydrogen production efficiency is established, and the deviation of the current temperature from the rated temperature is calculated to obtain the temperature coupling factor. The pressure influence factor is multiplied by the temperature coupling factor, and combined with the equipment performance attenuation value to obtain the mutual feedback correction parameter. The initial hydrogen production power regulation instruction is compensated and adjusted according to the mutual feedback correction parameter to obtain a corrected power instruction; at the same time, the maximum power limit of the electrolyzer is compensated to obtain a corrected power upper limit, and the power climbing rate limit of the electrolyzer is compensated to obtain a corrected climbing rate limit; A smoothing factor is determined based on the time rate of change of the mutual feed correction parameter, and the corrected power instruction and the power regulation instruction of the previous control cycle are weightedly averaged according to the smoothing factor to generate a transition instruction; it is judged whether the transition instruction meets the corrected power upper limit and the corrected climbing rate limit. If so, the transition instruction is output as the optimized hydrogen production power regulation instruction; if not, the transition instruction is limited within an allowable range and then output as the optimized hydrogen production power regulation instruction.

5. The method according to claim 1, wherein Monitor the operating status of the hydrogen production system. When an anomaly is detected, calculate the optimal power transfer path based on the fluctuation characteristic sequence, transfer the generated power to the grid according to the optimal power transfer path, and control the water electrolysis hydrogen production equipment to perform a step-by-step load reduction shutdown, including: The membrane voltage fluctuation rate, hydrogen production efficiency decay rate, and system pressure mutation rate of the water electrolysis hydrogen production equipment are collected to calculate the first-order rate of change and the second-order rate of change; when the first-order rate of change or the second-order rate of change exceeds the corresponding warning threshold, the power transfer mechanism is triggered; Obtaining a power generation sequence of a water electrolysis hydrogen production device and extracting power fluctuation characteristics; combining the first-order rate of change, the second-order rate of change, and the power fluctuation characteristics, and using a deep reinforcement learning algorithm to generate an initial power transfer path; Collect stress monitoring data of the water electrolysis hydrogen production equipment and calculate a stress state index based on the stress monitoring data; divide the load reduction process into a buffer load reduction section, a rapid load reduction section, and a protection load reduction section according to the stress state index; establish a dynamic mapping relationship between the stress state index and the load reduction rate of each load reduction section to generate a step-by-step load reduction strategy; Performing wavelet decomposition on the pressure signal of the hydrogen storage system to extract the disturbance component and the trend component; identifying the pressure pulsation characteristics based on the disturbance component, and predicting the pressure change trend based on the trend component; combining the pressure pulsation characteristics and the pressure change trend with the stress state index to obtain the system evolution characteristics; The system evolution characteristics are used to perform real-time correction on the initial power transfer path to obtain an optimal power transfer path; the generated power is controlled to be transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform load reduction based on the step-by-step load reduction strategy; The stress monitoring data and pressure signal feedback during the load reduction process are used to calculate the updated system evolution characteristics; based on the updated system evolution characteristics, the optimal power transfer path and the load reduction rate in the step-by-step load reduction strategy are optimized in real time; and the optimization process is repeated until the water electrolysis hydrogen production equipment is safely shut down.

6. The method according to claim 5, characterized in that Obtaining the power generation sequence of the water electrolysis hydrogen production equipment and extracting the power fluctuation characteristics; combining the first-order change rate, second-order change rate and power fluctuation characteristics, using a deep reinforcement learning algorithm to generate an initial power transfer path includes: Input power fluctuation characteristics into the strategy network and value network; The first hidden layer of the strategy network uses an exponential linear unit function to perform nonlinear combination on the power fluctuation characteristics to obtain a power feature vector, the second hidden layer processes the power feature vector through a residual connection to obtain a power adjustment action, and the output layer generates an action probability of power transfer based on the power adjustment action; The long-term and short-term memory units of the value network store power fluctuation characteristics to form a historical power sequence, and the importance of the historical power sequence is screened through a gating mechanism to obtain a state evaluation result; Calculating a power change acceleration coefficient, a power fluctuation amplitude coefficient, and a power response time coefficient according to the power fluctuation characteristics, constructing a power fluctuation constraint rule based on the power change acceleration coefficient, the power fluctuation amplitude coefficient, and the power response time coefficient, and combining the power fluctuation constraint rule with the action probability of the power transfer to form a multi-objective optimization function; The reward value is calculated using the multi-objective optimization function and the state evaluation result, a temporal difference error is constructed according to the reward value, a trust domain constraint is established based on the temporal difference error, the parameters of the policy network are optimized under the trust domain constraint, and an initial power transfer path is generated using the optimized policy network parameters.

7. The method according to claim 5, characterized in that Dividing the load shedding process into a buffer load shedding section, a rapid load shedding section, and a protection load shedding section according to the stress state index; Establishing a dynamic mapping relationship between the stress state index and the load shedding rate of each load shedding segment and generating a step-by-step load shedding strategy includes: Establishing a state characteristic matrix based on the stress state index and calculating a characteristic root combination of the state characteristic matrix; determining the system stability state based on the characteristic root combination and calculating a stability margin based on the system stability state; dividing the load shedding process into a buffer load shedding section, a rapid load shedding section, and a protection load shedding section based on the stability margin; Establishing a response function between load shedding rate and dynamic stress for each load shedding section; determining the initial load shedding rates of the buffer load shedding section, the rapid load shedding section, and the protection load shedding section based on the response function; A dynamic equation of stress state index and load reduction rate is established, wherein a nonlinear term is introduced to characterize the bifurcation characteristics of the system; the initial load reduction rate is substituted into the dynamic equation to solve the bifurcation parameter; Establishing an adaptive adjustment equation for stress deviation; calculating a compensation amount based on the deviation between the stress state indicator and a preset stress threshold; inputting the compensation amount into the adaptive adjustment equation to dynamically adjust the bifurcation parameter; solving the dynamic equation based on the adjusted bifurcation parameter to obtain the load shedding rate of each load shedding section; A neural network model with multiple hidden layers is constructed, historical load shedding data is input into the neural network model for training, a mapping relationship between load shedding rate and stress fluctuation is established using the trained neural network model, a load shedding rate correction value is calculated based on the mapping relationship, and the load shedding rate of each load shedding section is corrected in combination with the load shedding rate correction value; A load shedding time allocation function is established, and the switching moment and duration of each load shedding segment are calculated according to the time allocation function; and the switching moment, duration and the corrected load shedding rate are combined to obtain a step-by-step load shedding strategy.

8. A minute-level intelligent scheduling system for renewable energy hydrogen-based energy, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain real-time power generation data of renewable energy power generation equipment, operating parameter data of water electrolysis hydrogen production equipment, and hydrogen storage capacity data of the hydrogen storage system; The second unit is used to calculate the power generation fluctuation frequency based on the real-time power generation data, adaptively adjust the length of the dynamic time window based on the power generation fluctuation frequency, and use the dynamic time window to segment the power generation data to generate a fluctuation feature sequence; The third unit is used to build a minute-level intelligent scheduling model based on the fluctuation characteristic sequence, operating parameter data and hydrogen storage data, generate the initial hydrogen production power adjustment instruction, calculate the mutual feedback correction parameters between the hydrogen storage system and the water electrolysis hydrogen production equipment, and use the mutual feedback correction parameters to compensate and optimize the initial hydrogen production power adjustment instruction to obtain the optimized hydrogen production power adjustment instruction; The fourth unit is used to send the optimized hydrogen production power adjustment instruction to the control system of the water electrolysis hydrogen production equipment, control the operation of the water electrolysis hydrogen production equipment, and monitor the operating status of the hydrogen production system. When an abnormality is detected, the optimal power transfer path is calculated according to the fluctuation characteristic sequence, the generated power is transferred to the power grid according to the optimal power transfer path, and the water electrolysis hydrogen production equipment is controlled to perform a step-by-step load reduction and shutdown.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images